Pluggable embeddings.

Six providers in the box. Configure once in the admin SPA, then call /v1/{ns}/ingest_text and the server handles embed + store in one round trip.

The six providers

OpenAI

text-embedding-3-small (1536d)
text-embedding-3-large (3072d)
ada-002 (1536d, legacy)

Fields: api_key

Azure OpenAI

any deployment of text-embedding-3 family

Fields: endpoint · deployment · api_version · api_key

Google AI · Gemini

gemini-embedding-001 (current)
gemini-embedding-exp
text-embedding-004 (deprecated)

Fields: api_key

Voyage

voyage-3
voyage-3-lite
voyage-large-2
voyage-code-2

Fields: api_key

Cohere

embed-english-v3
embed-multilingual-v3
embed-english-light-v3

Fields: api_key

Ollama

nomic-embed-text
mxbai-embed-large
all-minilm

Fields: base_url (no key required)

Configure via the admin SPA

Settings → Embedding service

Open the admin SPA → Settings → Embedding service.
Pick a provider — the model dropdown updates with that provider's catalog.
Paste credentials. For Azure: also set endpoint, deployment, api_version.
Click Test — the server runs a real embed against the provider and reports OK or the error.
Click Save. Stored in-memory only; the API key is never echoed back.

If the configured model's dim differs from the namespace dim (default 768), the server pads with zeros or truncates automatically.

Configure via REST

PUT /v1/admin/embedding_config

curl -X PUT http://localhost:8000/v1/admin/embedding_config \
  -H "X-API-Key: $FEATHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "openai",
    "model":    "text-embedding-3-small",
    "api_key":  "sk-..."
  }'

Parameters

providerstrOne of openai, azure, gemini, voyage, cohere, ollama.
modelstrModel name from the provider's catalog (see GET /v1/admin/embedding_models).
api_keystrStored in-memory only. Never echoed back on subsequent GETs.
endpointstr?Azure only — https://<resource>.openai.azure.com
deploymentstr?Azure only — your deployment name.
api_versionstr?Azure only — e.g. 2024-02-01.
base_urlstr?Ollama only — e.g. http://localhost:11434.

One-call embed + store

POST /v1/{ns}/ingest_text

The headline endpoint. Server takes raw text, embeds it via the configured provider, and stores the vector + metadata in one round trip.

import requests

r = requests.post(
    "http://localhost:8000/v1/acme/ingest_text",
    headers={"X-API-Key": FEATHER_API_KEY},
    json={
        "text": "Summer 2026 Instagram ad — sandals, 30% off, women 25–40.",
        "metadata": {
            "namespace_id": "acme",
            "entity_id":    "creative_001",
            "attributes":   {
                "brand":    "acme",
                "channel":  "instagram",
                "campaign": "summer-2026",
            },
        },
    },
)
print(r.json())
# → {"id": 1729..., "namespace": "acme", "embedded": true, "dim": 768}

Inspect the catalog

GET /v1/admin/embedding_models

Returns the curated dropdown the admin SPA uses. Handy if you're building a custom settings UI.

curl -H "X-API-Key: $FEATHER_API_KEY" \
  http://localhost:8000/v1/admin/embedding_models

Atlas Admin SPA

REST API